feat(review): add NVIDIA NIM model provider

[seonghobae]

Why

The central OpenCode review pool exhausts public free models without a valid control conclusion. The organization now provides NVIDIA_NIM_API_KEY as a GitHub Actions secret.

Acceptance criteria

• Pass the secret only to the privileged central review job.
• Register NVIDIA NIM through the existing OpenCode provider seam.
• Add at least one current NVIDIA NIM chat model to the review candidate pool.
• Preserve fallback behavior when the secret is absent or rejected.
• Cover configuration and redaction behavior with the existing workflow/script tests.

Phase: Ops
Component: dot-github
Agent: Codex
Started: 2026-07-30T07:58:52Z

Related delivery: ContextualWisdomLab/saju-caldav#15

[seonghobae]

Status: the fix is ready but cannot self-merge — one admin action is required

Root cause confirmed from #672's failed run log: OPENCODE_MODEL_POOL_OUTCOME: exhausted (GitHub Models retirement, 2026-07-30). Because the required opencode-review check runs the base .github config under pull_request_target, no PR org-wide can currently earn approval evidence — including the PRs that fix the pool. #672 (this issue's NVIDIA NIM fix) is itself blocked for exactly this chicken-and-egg reason: its own opencode-review runs the still-exhausted base pool.

Two ready recovery PRs (both need one admin merge into .github main):

• feat(review): add NVIDIA NIM provider #672 — keyed NVIDIA NIM (nvidia/nemotron-3-ultra-550b-a55b, verified live on integrate.api.nvidia.com/v1), public-repo-gated, skip-when-secret-absent. Requires NVIDIA_NIM_API_KEY.
• feat(review): add keyless OmniRoute free-gateway provider to the OpenCode review pool #675 — keyless OmniRoute free-gateway provider (no per-provider secret): base URL from vars.OMNIROUTE_API_BASE_URL, optional key, public-repo-gated, skip-when-unconfigured. Complements feat(review): add NVIDIA NIM provider #672; useful if the NIM key is unavailable.

Admin recovery (only path — the scheduler cannot merge a PR whose required check is failing, and there is no non-merge override of the workflow-embedded model pool):

1. Bypass-merge feat(review): add keyless OmniRoute free-gateway provider to the OpenCode review pool #675 (keyless — also set vars.OMNIROUTE_API_BASE_URL to the org OmniRoute deployment) or feat(review): add NVIDIA NIM provider #672 (set/confirm NVIDIA_NIM_API_KEY) into .github main.
2. On merge, the central pool recovers org-wide; the scheduled org-queue-sweep then re-dispatches reviews and clears the frozen backlog.

For a keyless recovery to emit valid verdicts, the OmniRoute deployment must front at least one tool-calling-capable upstream (the reviewer needs real tool calls; raw reverse-engineered keyless upstreams alone can't).

---

Generated by Claude Code